AI Leaders Push Independent Oversight as Risks, Lobbying Rise

Anthropic and its rivals are flooding Washington with cash and dire warnings about artificial intelligence wiping out humanity, but the push for new regulatory guardrails looks more like an effort to consolidate power than genuine public protection.

Last week an Anthropic researcher publicly assigned greater than a 10 percent chance that advanced AI could kill all humans within the decade. The claim joined other recent alarms about synthetic pandemics and autonomous systems slipping human control. Microsoft AI chief Mustafa Suleyman added that keeping such systems aligned with humanity will prove a really big challenge. Dario Amodei, Anthropic’s CEO, has called for independent oversight bodies to monitor frontier models.

These statements coincide with aggressive spending. Anthropic routed $40 million through the super PAC Public First Action in recent months, including a $20 million check in July. The money has already flowed into congressional races, with $3 million directed toward one New York assembly candidate backed by the PAC. Lawmakers such as Sen. Ron Johnson have noted that the safety rhetoric could serve to limit competition and favor a handful of large, government-approved players.

The company’s board includes prominent political donors, and its lobbying footprint has expanded sharply. Rival firms have responded with their own PAC spending, turning AI policy into an expensive influence contest. Meanwhile, the technology’s actual near-term effects are already visible in cybersecurity. Microsoft alone patched 974 vulnerabilities this month. Oracle issued more than 1,400 fixes in July. Google Chrome and Mozilla reported record numbers of flaws uncovered with help from current AI tools. These problems strain security teams today, yet the public conversation centers on speculative future bioweapons or recursive self-improvement that lets models redesign themselves without human oversight.

Anthropic reports that its Claude model already handles 26 percent of the company’s own research and development tasks. Executives at several labs describe this capability as approaching autonomous improvement, though human supervision remains in place. The shift raises legitimate questions about long-term control. It also creates an opening for established players to argue that only they possess the resources to manage the risks, thereby justifying rules that raise barriers for newer entrants.

Critics inside and outside the industry point out that heavy regulation often entrenches incumbents who helped write the rules. The same pattern appeared in earlier technology sectors where safety or standards arguments produced concentrated markets. Public First Action has kept most of its recent funds in reserve, suggesting the spending campaign is still ramping up as Congress considers legislation.

The focus on extinction-level scenarios draws attention away from concrete, present-day issues such as the surge in software vulnerabilities and the concentration of advanced models in a few hands. If competition is the best safeguard against misuse, as Johnson and others argue, then policies that favor a small number of well-connected firms may increase rather than reduce exposure. The money now flowing into PACs and lobbying indicates the debate is already shaped as much by commercial interests as by technical warnings.